Two Years of LLMs in Cars — Why Has the True AI Vehicle Taken So Long?

Commanding a Car With a Single Sentence — This Time It's Real

At Auto China (Beijing Auto Show), a demo video of the Zeekr 8X went viral on social media.

In the video, a user sitting inside the car simply said: "Take me to pick up my kid from school, stop by a McDonald's on the way, and I need to be at the school before 5 o'clock." What happened next was fundamentally different from everything we've come to expect from a "smart car" — the vehicle autonomously planned a route, activated intelligent driving, pulled over at a McDonald's midway, and automatically parked upon arriving at the school entrance. Throughout the entire process, the user didn't need to touch the navigation, manually switch to autonomous driving mode, or hunt for a parking spot.

This wasn't a voice assistant helping you search for a restaurant. This was an Agent executing a task on your behalf.

Behind this seemingly simple interaction lies not just "a smart LLM plugged into a car," but a whole-vehicle intelligent agent system with full integration from "brain" to "limbs." The industry has been talking about putting AI in cars for three years — so why are we only now starting to see products like this?

From ChatBot to AI in Cars: A Smart Mouth, a Sluggish Body

Looking back at the past two years of the "LLMs in cars" craze, there's an uncomfortable truth — large models made it into vehicles, but never truly became part of the car.

From 2024 to 2025, virtually every automaker announced LLM integration. DeepSeek, Tongyi Qianwen, Doubao — they all took their turns on stage, and cabin voice assistants did get noticeably smarter. They could chat with you, look up encyclopedia entries, and the best ones had the entire vehicle manual loaded in, ready to tell you "what to do when the tire pressure warning comes on."

But at the end of the day, it was still just a chatbot running inside the cabin. The voice assistant stayed in its lane, navigation stayed in its lane, autonomous driving stayed in its lane, and parking stayed in its lane. They were like neighbors living in the same building who had never met — each minding their own business.

Whenever a user wanted to accomplish a slightly complex task — say, "find a highly-rated café, navigate there, and park for me when I arrive" — they had to play dispatcher themselves: first ask the voice assistant to search for a café, then manually enter the address into navigation, and then manually switch to parking mode upon arrival. The LLM's "intelligence" was trapped inside a chat window, completely severed from the vehicle's "body."

This was the real state of the industry over the past two years: a car with a smart mouth but clumsy limbs.

Why Is Full Integration So Hard? Three Technical Mountains

Upgrading an LLM from "able to chat" to "able to act" isn't a challenge of the model itself — it's a challenge of whole-vehicle architecture.

The first mountain: System decoupling. Traditional automotive electrical/electronic architectures are "siloed" — the cabin domain, autonomous driving domain, body domain, and powertrain domain each operate independently, developed by different suppliers, running different operating systems, communicating through limited signal buses. To let an AI brain simultaneously orchestrate navigation, autonomous driving, and parking, you first need to tear down these silos and establish a unified central computing platform. This isn't a software-level adaptation — it's a complete redesign of the vehicle's electronic architecture, a process that often takes two to three years.

The second mountain: From intent understanding to task decomposition. When a user says "be at the school by 5, and grab a McDonald's on the way," it's easy for a human to understand, but for a system, it requires a chain of complex reasoning: first understand the goal (pick up the child), constraints (before 5 PM), secondary tasks (McDonald's), and implicit needs (need to park), then decompose it into an executable sequence of sub-tasks — search for McDonald's, calculate the time window, plan the route, activate autonomous driving, make an intermediate stop, and finally park. This is a classic AI Agent task orchestration problem, requiring the LLM to have strong planning capabilities while dynamically binding to the vehicle's real-time state (location, battery level, traffic conditions).

The third mountain: Safety and determinism. LLMs are fundamentally probabilistic reasoning engines — they can hallucinate, they can make mistakes. On a phone, an AI assistant's error might mean recommending a bad restaurant; in a car, a wrong decision could be a matter of life and death. Therefore, a whole-vehicle intelligent agent cannot rely entirely on an LLM for end-to-end decision-making. It needs a hybrid architecture where "the LLM handles planning, and a rule engine provides guardrails." Which tasks can be delegated to AI for autonomous decision-making, and which must have hard-coded safety boundaries — where to draw this line is the core question every automaker is working to answer.

What Did Zeekr 8X Get Right?

Based on publicly available information, the Zeekr 8X was able to deliver this experience not by simply "plugging in an LLM," but by building a whole-vehicle-level intelligent agent system.

First, integration at the architecture level. Built on the SEA (Sustainable Experience Architecture) platform, Zeekr achieved central domain fusion across cabin, autonomous driving, and body control, giving the AI brain the "neural pathways" to orchestrate the entire vehicle. The LLM is no longer just an app inside the cabin — it has become the vehicle's "central decision-making layer."

Second, an Agent-based interaction paradigm. Users don't need to give precise commands — they only need to express intent. The system automatically handles intent parsing, task decomposition, subsystem orchestration, and execution feedback. This shift from "command-based interaction" to "intent-based interaction" is the fundamental difference between an AI vehicle and a traditional smart car.

Finally, end-to-end closed-loop capability. From understanding user needs, to route planning, to autonomous driving execution, to automatic parking — the entire chain requires no human intervention or manual switching. This means the system must not only "think it through" but also "follow through" — any break in the chain would reduce the experience back to square one.

Industry Inflection Point: From 'Model Arms Race' to 'Systems Integration Battle'

The emergence of the Zeekr 8X signals that AI vehicle competition is entering a new phase.

Over the past two years, automakers' AI narratives focused on "which LLM we've integrated" and "how smart our voice assistant is." This was essentially an arms race at the model level — low barrier to entry, limited differentiation — since everyone was plugging in the same handful of mainstream models, and converging experiences were only a matter of time.

What the Zeekr 8X demonstrates is a shift from "model capability" to "system capability." Whoever can truly infuse LLM intelligence into every execution layer of the vehicle will define the experience standard for the next generation of AI cars. This is no longer a problem a single algorithm team can solve — it requires full-stack coordination across vehicle architecture, software platforms, autonomous driving systems, and cloud services.

This also explains why the "true AI vehicle" has been so late to arrive. It wasn't waiting for a more powerful LLM — it was waiting for the entire automotive industry to complete the foundational shift from hardware-defined to software-defined, from distributed architecture to central computing. The LLM is the last "brain" placed on top, but without a torso and limbs, even the smartest brain is just a chatbot.

Outlook: Has the 'iPhone Moment' for AI Vehicles Arrived?

If we place the Zeekr 8X demo on a broader timeline, it is very likely a landmark moment marking the transition of AI vehicles from the "gimmick phase" to the "utility phase."

Here are several trends worth watching:

First, Agents will become the standard architecture for smart vehicles. It's not just Zeekr — NIO, XPeng, Huawei, and other players are also accelerating their whole-vehicle intelligent agent strategies. In the second half of 2025, we will likely see multiple automakers launch similar Agent-based products.

Second, evaluation criteria for AI vehicles will be redefined. In the past, a car's intelligence was measured by voice recognition accuracy, screen interaction smoothness, and how many scenarios the autonomous driving system could handle. In the future, the core metrics may become: How complex a task can be completed with a single sentence? How short is the chain from intent to execution?

Third, the data closed loop will become the core competitive moat. When AI truly starts making decisions and executing on behalf of users, every interaction generates valuable behavioral data — what needs users express in which scenarios, whether the system's planning was sound, whether the execution results met expectations. This data will drive continuous system evolution, creating a flywheel effect where the strong get stronger.

In the first two years of LLMs in cars, the industry completed the 0-to-1 "capability validation" phase. Starting with the Zeekr 8X, AI vehicles are officially entering the 1-to-100 "experience delivery" phase. The future where you sit in a car, say one sentence, and everything is taken care of is finally no longer just a vision on a PowerPoint slide.

It's hitting the road.